Event:

Skilled movements require the agent to learn to adapt to dynamic environments to improve performance over time. While many circuits from the cortex to the spinal cord are active and required for generating effective movements, it remains unclear how neural circuits across the brain coordinate, adapt, and contribute to movement. One limitation to studying movement in great detail was the ability to robustly measure the pose of an animal without using intrusive markers. In my talk, I will discuss a highly efficient method for markerless pose-estimation based on transfer learning with deep neural networks that achieves excellent results with minimal training data. After discussing various applications from odor guided navigation to cheetah hunting, I will focus on skilled reaching behavior in mice. Specifically, individual joints can be automatically tracked, and even when a small number of frames are labeled (~150), the algorithm achieves excellent tracking performance on test frames that is comparable to human accuracy. This is now allowing us to correlate neural activity across sensorimotor cortex with detailed 3D movements of the animal.